• Julio Saez-Rodriguez

Multi-omics technologies, specially with single-cell and spatial resolution, provide unique opportunities to the key study intra- and inter-cellular processes that drive immunological systems and their deregulation in disease. The use of prior biological knowledge allows us to reduce the dimensionality and increase the interpretability of the data, in particular by extracting from the data features describing the activity of molecular processes such as signaling pathways, gene regulatory networks, and cell-cell communication events. In this talk, I will present resources and methods that combine multi-omic single cell and spatial data with biological knowledge and illustrate them on medically relevant cases.

• Robert F. Murphy, Carnegie Mellon University, United States
• Huangqingbo Sun, Carnegie Mellon University, Stanford University, United States
• Anna Martinez Casals, Stanford University, United States
• Anna Bäckström, Stanford University, United States
• Yuxin Lu, Carnegie Mellon University, United States
• Cecilia Lindskog, KTH-Royal Institute of Technology, Sweden
• Matthew Ruffalo, Carnegie Mellon University, United States
• Emma Lundberg, Stanford University, United States

Identifying cell types in highly multiplexed images is essential for understanding tissue spatial organization. Current cell type annotation methods often rely on extensive reference images and manual adjustments. We have developed an open-source tool, Robust Image-Based Cell Annotator (RIBCA), that enables accurate, automated, unbiased, and fine-grained cell type annotation for multichannel tissue images without requiring additional model training or human intervention. It can be used with a wide range of antibody panels, initially focused on immune cell types. The design has two novel aspects. This first is an ensemble approach the merges a number of distinct deep learning models that each annotate different subsets of cell types using different sets of markers. That design allows for each extension to additional cell types without the need for retraining the existing models. The second is the use of auxiliary models to allow prediction of markers missing from a given panel. This provides much better assignment that replacing missing markers with a blank channel. Our tool has successfully annotated over 3 million cells, revealing the spatial organization of various cell types across more than 40 different human tissues.

• Haoyu Wang, University of Pittsburgh, United States
• Brittany Cody, University of Pittsburgh, United States
• Hatice Osmanbeyoglu, University of Pittsburgh, United States

Cell function in multicellular systems is shaped by its spatial context, making the study of cellular interactions within tissues essential for understanding development and disease. While spatial transcriptomic (ST) technologies capture genome-wide mRNA expression with spatial resolution, they do not provide protein-level insights, which are crucial for understanding cellular function and therapeutic targeting. Recent advancements like spatial CITE-seq enable simultaneous profiling of gene and protein expression. Here, we present a computational framework that imputes protein abundances from ST data by leveraging RNA–protein relationships learned from spatial CITE-seq using a Dual-Graph Attention Network (DGAT). Our method constructs heterogeneous graphs based on mRNA/protein expression and spatial coordinates, aligning these representations with Graph Attention Network encoders. These representations are decoded using a fully connected network for mRNA reconstruction and a multi-branch decoder for protein prediction. We applied DGAT to publicly available and in-house spatial CITE-seq datasets, including tonsil, breast cancer, glioblastoma, and malignant mesothelioma samples, demonstrating superior accuracy in imputing protein expression compared to methods that do not incorporate spatial information. We further applied DGAT to ST datasets from tonsil and breast cancer tissues, revealing deeper insights into cellular states, immune phenotypes, and spatial domains, such as germinal centers. Our approach enhances cell-type assignments, spatial domain detection, and the interpretation of ST datasets, advancing the understanding of tissue architecture and immune responses.

• Veronika Kedlian, Cambridge Stem Cell Institute, United Kingdom
• Lisa Marie Milchsack, Cambridge Stem Cell Institute, United Kingdom
• Marita Bosticardo, Laboratory of Clinical Immunology and Microbiology, Division of Intramural Research, NIAID NIH, United States
• Nadav Yayon, Cambridge Stem Cell Institute, United Kingdom
• Francesca Pala, Laboratory of Clinical Immunology and Microbiology, Division of Intramural Research, NIAID NIH, United States
• Luigi D. Notarangelo, Laboratory of Clinical Immunology and Microbiology, Division of Intramural Research, NIAID NIH, United States
• Sarah A. Teichmann, Cambridge Stem Cell Institute, United Kingdom

The thymus is a primary immune organ responsible for the production of naive T cells which experiences age-related changes, also termed involution very early in the lifespan. Despite the thymus crucial role in the adaptive immune system we lack a complete picture of the cell types and mechanisms involved in the process.
Here, we present the largest collection to date of thymus single-cell (~1M cells, 63 donors) and spatial sequencing data (Visium and Xenium spatial transcriptomics) across human lifespan and different stages of involution by integrating newly generated as well as publicly available datasets.
This resource has allowed us to identify and clarify changes in major cellular and spatial compartments of the thymus as well as their temporal order in the involution process. Broadly, we observe a loss in developing thymocytes and cTECs, and an expansion of the set of resident stromal populations in the adult thymus causing decline in T cell production. More specifically, we highlight two major bottlenecks in the T cell development which are implicated in the thymocyte decline. We also observe the expansion of dysfunctional mcTEC “progenitors” with age that is progressively biassed towards mTEC generation. In the stromal compartment, we identify putative fibroblast states implicated in adipose and fibrotic changes of the thymus.
Beyond these and other biological discoveries - we spatially position these populations using high-resolution spatial transcriptomics data, providing a valuable resource to study thymus and immune ageing for the community.

• Harinder Singh

• Alisa Omelchenko, University of Pittsburgh, United States
• Rebecca Elsner, University of Pittsburgh, United States
• Syed Rahman, University of Pittsburgh, United States
• Vinay Mahajan, Brigham and Women's Hospital, United States
• Jishnu Das, University of Pittsburgh, United States

Using networks approaches allows us to integrate multi-modal datasets and view the immune system pathways with a multi-scale lens. Therefore, designing methods which are interpretable are necessary for a holistic view of the immune system.
We developed a novel integrated network-framework to study T follicular helper (Tfh) cell differentiation. Tfh cell differentiation is a highly heterogeneous process that remains poorly understood and difficult to study due to experimental limitations. As a result, existing Tfh network diagrams are incomplete, with each study providing valuable but often disconnected information. Specifically, we identified subnetworks by integrating epigenomic or transcriptomic signals with a protein-protein interaction (PPI) network through network propagation. We also detected key transcription factors (TFs) by merging epigenomic signals with the transcriptional regulatory network using Personalized Page Rank.
In addition to capturing well-known circuits, we unveiled novel modules integral to Tfh cell differentiation including the IL-12/IL-23/IL-27 pathway and SHP2 signaling. IL-12 has been a controversial element, and its role is highly debated. Our network shows in an unbiased manner the function of IL-12 in the network. Further, while the functions of Tfh cells and their cooperative roles with B cells are broadly similar in mice and humans, several differences have previously been reported which have also been contested. We underscore significant similarities between human and murine Tfh networks, suggesting a higher degree of conservation than previously reported. These insights provide a more cohesive understanding of the regulatory mechanisms governing Tfh cell differentiation and pave the way for therapeutic interventions targeting humoral immunity.

• Isabella Withnell, University College London, United Kingdom
• Zara Baig, University College London, United Kingdom
• Joseph Ng, University College London, United Kingdom
• Franca Fraternali, University College London, United Kingdom
• Claudia Mauri, University College London, United Kingdom

Pan cancer atlases often integrate cells across tissues, obscuring context specific immune states. We developed a tissue first single cell RNA seq pipeline that (i) clusters B cells within each cancer type and then (ii) embeds the resulting clusters into a shared latent space learned by a variational auto encoder (VAE). This approach preserves rare phenotypes and enables quantitative, inter cluster distance measurements.

Across four tumour types (breast, colorectal, lung, renal; 97,450 B cells), our approach separated tumour conserved from tissue restricted programs. Renal tumours were enriched for a Double Negative 1 (DN1) B cell subset (CD27⁻ IGHD⁻ CR2⁺) with high IL 10, IL 23A, TIGIT and hypoxia/osmotic stress signatures, suggesting environment driven expansion. Spatial immunohistochemistry and flow cytometry confirmed enrichment of these cells in the renal tumour, where they are in dysfunctional tertiary lymphoid structures. They suppressed CD8⁺ T cell cytotoxicity in vitro and enrichment is correlated with poor OS.

Our framework recovers immune diversity hidden by global integrations and uncovers a kidney enriched B cell population with therapeutic relevance.

• Dongjun Guo, University College London, United Kingdom
• Deborah Dunn-Walters, University of Surrey, United Kingdom
• Franca Fraternali, University College London, United Kingdom
• Joseph Ng, University College London, United Kingdom

The development of stable antibodies formed by compatible heavy (H) and light (L) chain pairs is crucial in both the in vivo maturation of antibody-producing cells and the ex vivo designs of therapeutic antibodies. We present here a novel machine learning framework, ImmunoMatch, for deciphering the molecular rules governing the pairing of antibody chains. Fine-tuned on an antibody-specific language model, ImmunoMatch learns from paired H and L sequences from single human B cells to distinguish cognate H-L pairs and randomly paired sequences. We find that the predictive performance of ImmunoMatch can be augmented by training separate models on the two types of antibody L chains in humans, κ and λ, in line with the in vivo mechanism of B cell development in the bone marrow. Using ImmunoMatch, we illustrate that refinement of H-L chain pairing is a hallmark of B cell maturation in both healthy and disease conditions. We find further that ImmunoMatch is sensitive to sequence differences at the H-L interface. ImmunoMatch focusses on H-L chain pairing as a specific, under-explored problem in antibody developability, and facilitates the computational assessment and modelling of stably assembled immunoglobulins towards large-scale optimisation of efficacious antibody therapeutics.

• Lindeberg Leite, Department of Computer Science, University of Brasilia, Brazil
• Teófilo de Campos, Applied Sciences Group, Microsoft UK, United Kingdom
• Francisco Lobo, National Cancer Institute, United States
• Felipe Campelo, Intelligent Systems Laboratory, University of Bristol, United Kingdom

The identification of linear B-cell epitopes (LBCEs) is an important step in the development of immunodiagnostic tests and vaccines. Most existing computational methods for LBCE prediction are generalist models and do not incorporate explicit information on the target pathogen or its evolutionary relationships with other organisms present in the training data. This can lead to biases toward well-studied pathogens or taxa, with poorer performance for emerging or neglected infectious agents. To address this limitation, we present a phylogeny-aware framework that enhances LBCE prediction by incorporating evolutionary relationships into model training. Our approach employs taxonomy as a proxy to phylogeny and uses it to curate the training data. We introduce a transfer learning strategy that fine-tunes large protein language models using data available for higher-level taxa before deploying them to create a pathogen- or taxon-specific predictive model. This phylogeny-aware feature embedding substantially improves predictive accuracy compared to state-of-the-art methods, particularly but not exclusively for data-scarce or understudied pathogens. By leveraging evolutionary relationships, our framework optimises the use of available epitope data and provides more accurate LBCE prediction for emerging or neglected infectious agents. We report computational results for 20 target taxa including viral, bacterial and eukaryotic pathogens, which indicate median AUC gains between 0.15 and 0.2 in relation to current methods. Reference: The results presented in this work are described in greater detail in our paper "EpitopeTransfer: a Phylogeny-aware transfer learning framework for taxon-specific linear B-cell epitope prediction", currently under review.

• Pengfei Zhang, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States
• Hao Mei, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States
• Seojin Bang, Google DeepMind, United States
• Heewook Lee, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States

Reliable TCR-epitope binding prediction models are essential for development of adoptive T cell therapy and vaccine design. These models often struggle with false positives, which can be attributed to the limited data coverage in existing negative sample datasets. Common strategies for generating negative samples, such as pairing with background TCRs or shuffling within pairs, fail to account for model-specific vulnerabilities or biologically implausible sequences. To address these challenges, we propose an iterative attack-and-defend framework that systematically identifies and mitigates weaknesses in TCR-epitope prediction models. During the attack phase, a Reinforcement Learning from AI Feedback (RLAIF) framework is used to attack a prediction model by generating biologically implausible sequences that can easily deceive the model. During the defense phase, these identified false positives are incorporated into fine-tuning dataset, enhancing the model's ability to detect false positives. A comprehensive negative control dataset can be obtained by iteratively attacking and defending the model. This dataset can be directly used to improve model robustness, eliminating the need for users to conduct their own attack-and-defend cycles. We apply our framework to five existing binding prediction models, spanning diverse architectures and embedding strategies to show its efficacy. Experimental results show that our approach significantly improves these models' ability to detect adversarial false positives. The combined dataset constructed from these experiments also provides a benchmarking tool to evaluate and refine prediction models.

• Ko-Han Lee, University of California San Diego, United States
• Timothy Sears, University of California San Diego, United States
• Maurizio Zanetti, University of California San Diego, United States
• Hannah Carter, University of California San Diego, United States

Cancer immunotherapy, including immune checkpoint inhibitors (ICIs) and personalized cancer vaccines, has transformed cancer treatment. However, response rates remain suboptimal, highlighting the need for more effective strategies. Accurate identification of immunogenic neoantigens is critical to improving therapeutic outcomes, yet current prediction methods have key limitations—including insufficient modeling of T-cell recognition, limited incorporation of major histocompatibility complex (MHC) class II pathways, and a lack of tumor clonality integration.

To address these challenges, we developed NeoPrecis, a computational framework that combines enhanced T-cell recognition modeling with comprehensive immunogenicity and tumor clonality analysis. It comprises two modules: NeoPrecis-Immuno, which predicts T-cell recognition by modeling cross-reactivity distances between wild-type and mutant peptides, and NeoPrecis-Landscape, which integrates immunogenicity predictions with clonal architecture to derive a tumor-centric immunogenicity score.

NeoPrecis-Immuno achieves superior accuracy on an independent gastrointestinal cancer dataset with validated CD4+ and CD8+ T-cell assays, outperforming state-of-the-art predictors including PRIME, DeepNeo, and ICERFIRE. The MHC-binding motif enrichment step enhances the model’s ability to capture features relevant to peptide-MHC (pMHC) and T-cell receptor (TCR) interaction. A derived motif benefit score—even in the absence of specific mutations—shows significant association with patient survival in melanoma (p = 0.04) and non-small cell lung cancer (NSCLC, p = 0.01). NeoPrecis-Landscape further outperforms tumor mutation burden (TMB) in predicting ICI response, particularly in melanoma and in heterogeneous NSCLC tumors with low immunoediting.

Together, these findings highlight NeoPrecis as a robust, interpretable tool for neoantigen prioritization and personalized cancer immunotherapy guidance.

• Antonio Collesei, Veneto Oncology Institute (IOV-IRCSS), Italy
• Francesco Spinnato, University of Pisa, Italy
• Pierangela Palmerini, Veneto Oncology Institute (IOV-IRCSS), Italy
• Emilia Vigolo, Veneto Oncology Institute (IOV-IRCSS), Italy

Recent advances in RNA sequencing technologies and the gradual decrease in costs have allowed to design serial experiments with time embeddings, even at single cell resolution (scRNAseq). This possibility unlocks a finer level of detail, as well as a huge amount of noisy information to decode. Tools inferring regulatory networks, or patterns, from this type of data often focus on trajectories, disregarding local shapes and fundamental time series primitives. Moreover, they fail to target the analysis on a few meaningful results, reporting large, noisy outputs that need further downstream analysis, especially considering the intricate protein interplay happening within, for example, the immune context.

We describe SHISMA, a novel tool to infer significant cell type-specific regulatory patterns, or subnetworks, with strong statistical guarantees in terms of p-value. SHISMA exploits a time series primitive, the Bag-of-Patterns, adapted to discretize shorter temporal data (that is, with few time embeddings) and retain local shapes. SHISMA extracts significant patterns by performing a random walk approach on a protein-protein interaction network, with nodes identified by genes and scores derived from the shape-induced representation of the data, while properly validating via permutation and correcting for multiple hypothesis testing.

Our extensive experimental evaluation on synthetic data shows that our tool is able to retrieve specific and significant subnetworks from single cell time series transcriptomic data. Moreover, being tested on a real-world B-cell scRNAseq dataset, the subnetworks identified by SHISMA confirm its ability to retrieve known cell type-specific immunological processes, as well as potentially novel patterns and regulatory mechanisms.

• Nico Henschel, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, Germany
• Vaishnavee Ms. Thote, University of Bonn, Germany, Systems Medicine, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, Germany
• Pia Grundschoettel, LIMES Institute, University of Bonn, Germany, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, Germany
• Thomas Ulas, LIMES Institute, University of Bonn, Germany, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, Germany
• Divya Sitani, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, Germany
• Joachim L. Schultze, LIMES Institute, University of Bonn, German Center for Neurodegenerative Diseases (DZNE), Bonn, University of Bonn, Germany

Early life vaccination with whole cell or acellular pertussis vaccines shapes long term immune trajectories that influence responses to booster immunizations. In this study, we analyzed immune responses following tetanus, diphtheria, and acellular pertussis booster vaccination to investigate how infancy vaccination influences recall responses later in life. We applied machine learning to immune data from the CMI-PB Challenge, including gene expression, antibody titers, cytokine levels, and cell frequencies from annual donor cohorts collected from 2020 to 2022.

Each measurement type was treated as a separate modality. We applied cohort wise normalization and SHAP based feature selection within each modality, followed by feature level fusion to integrate selected features across modalities. A range of classifiers, including random forests, SVM, KNN, logistic regression, multilayer perceptrons, and XGBoost was applied on individual modalities, pairs, and fused datasets to distinguish between whole cell and acellular priming.

SHAP analysis identified IgG4 antibodies to filamentous hemagglutinin and pertussis toxin, along with cytokines such as CCL8, CCL2, IL1 alpha, and CXCL9, as key predictors, suggesting that repeated boosting may shape both antibody profiles and cytokine driven immune responses. In model evaluations, no individual modality consistently outperformed others across cohorts. For example, training on 2021 and 2022 and testing on 2020, gene expression achieved AUROC 0.859 while multimodal model reached 0.958. When testing on 2022, antibody features yielded AUROC 0.792 and the multimodal model achieved 0.866, potentially reflecting immune signatures shaped by COVID-19 vaccination. These findings highlight the importance of multimodal fusion for cohort generalizable immune response prediction.

• Romina Appierdo, Department of Biology, University of Rome “Tor Vergata”, Rome, Italy, Italy
• Pier Federico Gherardini, Department of Biology, University of Rome “Tor Vergata”, Rome, Italy, Italy
• Francesco Vallania, Machine Learning Collective, San Francisco, CA, United States
• Marina Sirota, Bakar Computational Health Sciences Institute, University of California San Francisco; San Francisco CA, USA, United States
• Manuela Helmer-Citterich, Department of Biology, University of Rome “Tor Vergata”, Rome, Italy, Italy
• Gerardo Pepe, Department of Biology, University of Rome “Tor Vergata”, Rome, Italy, Italy

Autoimmune diseases are heterogeneous and multifactorial, making it challenging to identify unifying mechanisms or clinically useful biomarkers. Despite abundant transcriptomic data, there is no integrated framework that systematically captures shared and tissue-specific immune dysregulation across diseases.

Here, we curated a large-scale transcriptomic compendium of 13,263 samples across 10 autoimmune diseases, integrating 6,238 blood and 7,025 tissue profiles from 156 public studies. Using meta-analysis, we derived robust, reproducible gene signatures for each disease, providing a foundational resource for downstream applications such as biomarker discovery and computational drug repurposing.

To enhance interpretability, we used computational strategies to extract 700 immune-relevant features—including pathway activity, cytokine expression, immune cell proportions, and regulatory signatures from transcription factors and miRNAs. We organized these features into 15 biologically coherent modules based on correlation patterns across diseases. Each module summarizes key immunological processes, such as interferon responses or adaptive immune triggering, enabling systematic and interpretable comparisons across diseases.

Subsequently, to address a major gap in the field, we systematically compared immune activity in blood and tissue across diseases. This revealed modest gene-level overlap but strong compartment-spanning coordination for specific pathways, particularly interferon signaling.

Finally, we demonstrated the translational relevance of this modular framework through multiple clinical applications. Modules predicted treatment response in inflammatory bowel disease (AUC up to 0.80), correlated with clinical severity in psoriasis (R > 0.7 with PASI score), and differentiated responders from non-responders before treatment. Modules also responded as expected to immune perturbation in controlled stimulation experiments, confirming biological validity.

• Bünyamin Şen, Hacettepe University, Turkey
• Tunca Doğan, Hacettepe University, Turkey

Accurate and reliable prediction of B cell epitopes holds critical importance in immunology and vaccine development. While traditional experimental methods offer high accuracy in identifying epitope regions, they are often laborious, time-consuming, and costly. Therefore, attempts are made to increase the efficiency of experimental characterization processes by using computational approaches. Since approximately 90% of epitopes are conformational, the prediction processes must account for three-dimensional protein structures and the geometric details of antigen-antibody interactions. In response to these requirements, our study introduces BepiCon (B-cell EPİtope Prediction Using CONtrastive learning), a two-step geometric deep learning framework that models antigen proteins as graph structures, incorporating structural and physicochemical properties and protein language model embeddings to automatically predict epitope regions on antigen proteins. In the first stage, the model was trained using a graph contrastive learning approach to learn high-quality representations of epitope and non-epitope residues. In the second stage, the pre-trained model was fine-tuned using supervised learning to perform conformational epitope prediction. The developed framework has demonstrated effective and generalizable performance when applied to both experimentally determined protein structures and predicted structures. Comparative analysis revealed that our approach distinguishes itself from existing B cell epitope prediction methods by exhibiting a lower false-positive rate and generating more reliable predictions. Our work contributes significantly to scientific research and therapeutic design processes by showcasing the advantages of geometric deep-learning approaches in B-cell epitope prediction.

• Jonas Schuck, Goethe University, University Hospital Frankfurt, Institute of Neurology (Edinger Institute) Frankfurt/Main, Germany
• Samira Ortega Iannazzo, Goethe University, University Hospital Frankfurt, Institute of Neurology (Edinger Institute) Frankfurt/Main, Germany
• Zeina Yasser Mahmoud, Goethe University, University Hospital Frankfurt, Institute of Neurology (Edinger Institute) Frankfurt/Main, Germany
• Lucie Marie Hasse, Goethe University, University Hospital Frankfurt, Institute of Neurology (Edinger Institute) Frankfurt/Main, Germany
• Katharina Imkeller, Goethe University, University Hospital Frankfurt, Institute of Neurology (Edinger Institute) Frankfurt/Main, Germany

Characterizing the antigen receptor repertoire of adaptive immune cells in solid tumors is essential for understanding the dynamics of immune responses across diverse cancer types. Adaptive immune receptor repertoire (AIRR) studies commonly rely on short-read sequencing, primarily capturing complementarity-determining region 3 (CDR3) sequences. However, full-length immunoglobulin and T cell receptor sequences offer richer structural information, including allelic variants, somatic hypermutations, and constant regions that define receptor isotypes. Transcripts encoding B and T cell receptors in 10x Genomics Visium-based spatial transcriptomics experiments range from 500 to 3500 base pairs, which requires long-read sequencing to resolve their full sequence architecture. While Oxford Nanopore Technologies (ONT) offers a cost-effective solution for long-read sequencing, robust bioinformatic tools for processing and annotating full-length antigen receptor sequences are lacking.

To fill this gap, we introduce NanoAIRR. This bioinformatic toolbox combines established software with novel functionality to enable accurate annotation of productive full-length immunoglobulin and T cell receptors at spatial tissue resolution. NanoAIRR is implemented as a modular bash tool comprising six main functionalities. These can be used independently or integrated into a streamlined end-to-end pipeline, e.g. by utilizing Snakemake. By integrating ultra-high-accuracy basecalling and error correction through Unique Molecular Identifiers, we improve long-read sequencing quality, enhancing the reliability of the downstream analysis.

We demonstrate how NanoAIRR enables us to study the spatial organization of adaptive immune receptors, track clonal lineage relationships, and assess the presence and structure of tertiary lymphoid structures (TLS) across various tumor types.